Clinical bacterial infection factor prediction model for lung transplantation patient and construction method

Abstract

The invention relates to the technical field of medical information, and discloses a lung transplantation patient clinical bacterial infection factor prediction model and a construction method, and the method comprises the following steps: S1, selecting a hospital needing research, and determining a research object according to an inclusion standard and an exclusion standard; s2, selecting a proper number of variables as candidate predictive factors; s3, clinical bacterial infection patients are defined as typical symptoms, signs and sputum characters; laboratory inspection: inflammation indexes are increased, and pathogens can be determined by bacteria detected by sputum culture; s4, if missing values exist in part of variables, simple interpolation processing is adopted, mean interpolation is adopted for continuous variables, and mode interpolation is adopted for classified variables; s5, sample size analysis; and S6, model establishment: firstly carrying out description analysis, and then carrying out Lasso regression and model prediction. A method for respectively establishing prediction models by applying various machine learning and deep learning is adopted, and multi-aspect comparative analysis is carried out on lung transplantation patient clinical bacterial infection influence factor prediction models.

Description

Technical Field

[0001] The present invention relates to the field of medical information technology, and in particular to a prediction model for clinical bacterial infection factors in lung transplant patients and a construction method thereof. Background Art

[0002] Lung transplantation has become the standard treatment for various non-malignant causes of advanced lung diseases, but the median survival of lung transplant recipients is about 6 years, which is much lower than that of other solid organ transplant recipients. Infection is an important cause of short-term death in patients after lung transplantation. It is the second leading cause of death within 30 days after surgery and the first leading cause of death within 1 year after surgery. Studies have shown that bacterial microorganisms have been isolated in nearly 80% of lung transplant recipients. Therefore, we need tools to detect infected patients early and take preventive measures in time to improve their quality of life.

[0003] Current studies have shown that bacterial infection of the donor lung and bacterial infection of the recipient's lung before surgery are risk factors for early bacterial lung infection after lung transplantation, but it is difficult to predict a patient's risk based on a single factor, and a comprehensive tool is needed to help determine it. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a prediction model for clinical bacterial infection factors in lung transplant patients and a construction method thereof, which solves the problem of difficulty in predicting bacterial infection in patients after lung transplantation.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients, comprising the following steps:

[0006] S1: Determine the research subjects: Select the hospitals to be studied and determine the research subjects according to the inclusion and exclusion criteria;

[0007] S2: Selection of candidate predictors: Select an appropriate number of variables as candidate predictors;

[0008] S3: Definition of outcome indicators: Patients with clinical bacterial infection are defined as: typical symptoms, signs and sputum characteristics; laboratory examination: elevated inflammatory indicators, sputum culture detection of bacteria to identify pathogens; imaging examination: pneumonia chest CT or X-ray examination shows new or progressive infiltration, consolidation or ground-glass shadows, bronchitis imaging examination is normal or chest X-ray shows moderate interstitial exudate;

[0009] S4: Missing value processing: Some variables have missing values, which are processed by simple interpolation: continuous variables use mean interpolation, and categorical variables use mode interpolation;

[0010] S5: Sample size analysis: First, the initial data set is obtained. When the ratio of the infected group and the non-infected group of the variable "clinical bacterial infection" is greatly deviated, the SMOTE upsampling algorithm is used to adjust the ratio of the two groups of variables to a certain ratio, and finally the sample size is obtained;

[0011] S6: Model building: First, descriptive analysis is performed, followed by Lasso regression and prediction model.

[0012] Preferably, the inclusion criteria in step S1 include the selection of lung transplantation surgery and unilateral or bilateral lung transplantation within a certain period of time, and the exclusion criteria include multi-organ transplantation, secondary transplantation, recipient age <18 years old and lack of complete medical history data.

[0013] Preferably, the increased inflammatory indicators in step S3 include increased white blood cells, neutrophils, C-reactive protein and serum procalcitonin.

[0014] Preferably, the analysis described in step S6 includes expressing the measurement data that obeys normal distribution as mean ± standard deviation, and comparing between groups using independent sample t test;

[0015] The measurement data that did not conform to normal distribution were expressed as median and interquartile range, and the MannWhitney U test was used for comparison between groups;

[0016] Enumeration data were expressed as number of cases and composition ratio. Pearson χ2 test or Fisher's exact probability method were used for comparison between groups of unordered categorical data, and MannWhitney U test was used for comparison between groups of ordered categorical data.

[0017] Univariate analysis showed that 17 variables including age at diagnosis, BMI, and WHO had p values ​​less than 0.1, and these variables were included in subsequent model analyses.

[0018] Preferably, the LASSO regression in step S6 is a linear regression method for feature selection and parameter estimation by adding an L1 regularization term, and its formula is:

[0019] minimize1 / 2*||Y-Xβ||^2+λ*||β||

[0020] Among them, Y is the dependent variable vector, X is the independent variable matrix, β is the regression coefficient vector, ||.|| represents the norm, ||Y-Xβ||^2 represents the residual sum of squares, that is, the sum of the squares of the differences between the predicted values ​​and the actual observed values, and ||β|| is the L1 norm, which represents the sum of the absolute values ​​of the regression coefficient vector.

[0021] Preferably, the prediction model in step S6 includes logistic regression, random forest, deep neural network and convolutional neural network;

[0022] The Logistic regression randomly divides the data set into a training set and a test set in a ratio of 7:3, and incorporates the LASSO regression screening variables into the l2 regularized Logistic. The l2 regularization algorithm reduces the complexity and overfitting risk of the Logistic regression model by adding a square sum of a parameter multiplied by a regularization coefficient in the loss function, thereby ultimately determining the optimal model parameters, where the l2 regularization coefficient is 1.0.

[0023] Preferably, the random forest uses the same training data set as Logistic regression, and the variables included are also Lasso regression screening variables. The variable selection standard adopts the Gini coefficient and then undergoes 5-fold cross validation to finally determine the optimal model parameters.

[0024] Preferably, the deep neural network uses the same training data set as Logistic regression, and the variables included are also Lasso regression screening variables, and then uses a fully connected neural network, the activation function uses the LeakyReLU function, the optimizer uses the Adama strategy, and the cross entropy loss is used as the model optimization target to finally determine the optimal model parameters.

[0025] Preferably, the convolutional neural network uses the same training data set as Logistic regression, and the variables included are also Lasso regression screening variables. Then, through the neural network, the activation function uses the LeakyReLU function, the optimizer uses the Adama strategy, and the cross entropy loss is used as the model optimization target to finally determine the optimal model parameters.

[0026] The present invention provides a prediction model for clinical bacterial infection factors in lung transplant patients and a construction method thereof.

[0027] It has the following beneficial effects:

[0028] 1. The present invention uses a variety of machine learning and deep learning methods to establish prediction models, and conducts a multi-faceted comparative analysis of the prediction model of clinical bacterial infection influencing factors in lung transplant patients. The model can identify high-risk patients early, take preventive and treatment measures in time, effectively reduce postoperative infection rates, and improve patients' survival rate and quality of life.

[0029] 2. The present invention adopts Lasso regression for feature selection, combined with a variety of advanced machine learning and deep learning algorithms such as L2 regularized Logistic regression, random forest, deep neural network and convolutional neural network to ensure that the model has high predictive performance and stability.

[0030] 3. Through comparative analysis of multiple performance indicators (such as AUC, Brier score, decision curve, etc.), the optimal model (convolutional neural network) selected in the present invention has excellent performance in terms of sensitivity, specificity, accuracy, etc., and has good classification ability and prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flow chart of a method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients according to the present invention;

[0032] Figure 2 It is the LASSO regression coefficient distribution curve diagram of the present invention;

[0033] Figure 3 It is the LASSO regression lambda decision diagram of the present invention;

[0034] Figure 4 Schematic diagram of the importance of variables in the random forest model of the present invention;

[0035] Figure 5 The ROC curve diagram of the prediction model of the influencing factors of clinical bacterial infection in four lung transplant patients of the present invention;

[0036] Figure 6 It is a decision curve diagram of the prediction model of influencing factors of clinical bacterial infection in four lung transplant patients of the present invention;

[0037] Figure 7 The training set ROC curve diagram of the Delong test of the prediction model of the influencing factors of clinical bacterial infection in four lung transplant patients of the present invention;

[0038] Figure 8 It is the test set ROC curve diagram of the Delong test of the prediction model of the influencing factors of clinical bacterial infection in four lung transplant patients of the present invention. DETAILED DESCRIPTION

[0039] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] Please see attached Figure 1 -Attached Figure 8 The present invention provides a method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients, comprising the following steps:

[0041] S1: Determine the research subjects: Select the hospitals to be studied and determine the research subjects according to the inclusion and exclusion criteria;

[0042] S2: Selection of candidate predictors: Select an appropriate number of variables as candidate predictors;

[0043] S3: Definition of outcome indicators: Patients with clinical bacterial infection are defined as: typical symptoms, signs and sputum characteristics; laboratory examination: elevated inflammatory indicators, sputum culture detection of bacteria to identify pathogens; imaging examination: pneumonia chest CT or X-ray examination shows new or progressive infiltration, consolidation or ground-glass shadows, bronchitis imaging examination is normal or chest X-ray shows moderate interstitial exudate;

[0044] S4: Missing value processing: Some variables have missing values, which are processed by simple interpolation: continuous variables use mean interpolation, and categorical variables use mode interpolation;

[0045] S5: Sample size analysis: First, the initial data set is obtained. When the ratio of the infected group and the non-infected group of the variable "clinical bacterial infection" is greatly deviated, the SMOTE upsampling algorithm is used to adjust the ratio of the two groups of variables to a certain ratio, and finally the sample size is obtained;

[0046] S6: Model building: First, descriptive analysis is performed, followed by Lasso regression and prediction model.

[0047] The specific steps are as follows:

[0048] 1. Determine the research subject

[0049] The data were obtained from hospitals that needed to be studied. The inclusion criteria were as follows: (1) lung transplantation between January 2018 and December 2022; (2) unilateral or bilateral lung transplantation. The exclusion criteria were as follows: (1) multi-organ transplantation; (2) secondary transplantation; (3) recipient age <18 years; (4) lack of complete medical records.

[0050] The sample size was 717 cases.

[0051] 2. Selection of candidate predictors

[0052] Thirty-seven variables were selected as candidate predictors (donor type, donor age, donor gender, donor BMI, donor ABO blood type, cold ischemia time, blood type matching, recipient age (years), recipient gender, recipient ethnicity, recipient BMI, recipient ABO blood type, major lung disease diagnosis classification, smoking status, pneumothorax and bullae, number of preoperative pulmonary complications, pulmonary embolism, pulmonary infection, pulmonary sepsis requiring treatment 2 or more times in the past year, pan-resistant bacterial infection, hormone dependence, cor pulmonale, coronary atherosclerotic heart disease, number of preoperative cardiac and cerebrovascular complications, functional status, liver dysfunction, renal dysfunction, hypokalemia, dialysis, diabetes, hypertension, surgical history, tumor history, oxygenation index, transplant type, operation time, intraoperative ECMO type).

[0053] 3. Define outcome measures

[0054] Lung transplant patients with clinical bacterial infection are defined as: ① typical symptoms, signs and sputum characteristics; ② laboratory examination: elevated inflammatory indicators (such as white blood cells, neutrophils, C-reactive protein and serum procalcitonin, etc.), and bacteria detected by sputum culture can identify the pathogen; ③ imaging examination: new or progressive infiltration, consolidation or ground-glass shadows on chest CT or X-ray for pneumonia, and normal imaging examination or moderate interstitial exudate on chest X-ray for bronchitis.

[0055] 4. Missing Value Handling

[0056] Some variables have missing values, which are processed by simple interpolation, that is, continuous variables are interpolated by mean value, and categorical variables are interpolated by mode value.

[0057] 5. Sample size analysis

[0058] The initial data set contained 717 records. The ratio of the infected group and the non-infected group of the dependent variable "clinical bacterial infection" was 591 / 126, which had a large deviation. The smote upsampling algorithm was used to sample the ratio of the two groups of variables to a ratio of 591 / 592, resulting in a final sample size of 1,183 cases.

[0059] 6. Model building

[0060] 6.1 Descriptive Analysis

[0061] Measurement data that followed normal distribution were expressed as mean ± standard deviation (Mean ± SD), and independent sample t test was used for inter-group comparison. Measurement data that did not follow normal distribution were expressed as median and interquartile range [M(Q1, Q3)], and Mann-Whitney U test was used for inter-group comparison. Count data were expressed as number of cases and constituent ratio [n(%)]. Pearson χ2 test or Fisher's exact probability method was used for inter-group comparison of unordered categorical data, and Mann-Whitney U test was used for inter-group comparison of ordered categorical data.

[0062] Univariate analysis showed that 17 variables including diagnosis age, BMI and WHO had p values ​​less than 0.1, and these variables were included in the subsequent model analysis.

[0063] 6.2 Lasso Regression

[0064] LASSO (Least Absolute Shrinkage and Selection Operator) regression is a linear regression method that performs feature selection and parameter estimation by adding L1 regularization terms. Its formula is:

[0065] minimize1 / 2 * ||Y-Xβ||^2+λ*||β||

[0066] Among them, Y is the dependent variable vector; X is the independent variable matrix; β is the regression coefficient vector; ||.|| represents the norm; ||Y-Xβ||^2 represents the residual sum of squares, that is, the sum of the squares of the differences between the predicted value and the actual observed value; ||β|| is the L1 norm, which represents the sum of the absolute values ​​of the regression coefficient vector; λ is the regularization parameter, which is used to weigh the importance between the residual term and the regularization term. LASSO regression finds the optimal regression coefficient β by minimizing the above formula. Due to the existence of the L1 regularization term, LASSO regression has sparsity, which can compress the regression coefficients of unimportant features to 0, thereby achieving the effect of feature selection. By adjusting the size of λ, the sparsity and the complexity of the model can be controlled.

[0067] Taking clinical bacterial infection as the dependent variable, LASSO regression was used to screen the independent variables. Through 10-fold cross validation, the optimal lambda was finally determined to be 0.019, and 16 independent variables were screened out, namely: donor type, donor BMI, donor ABO blood type, cold ischemia time (h), major lung disease diagnosis classification, recipient age (years), recipient BMI, smoking, pulmonary embolism, lung infection, pulmonary sepsis requiring treatment 2 or more times in the past year, pan-resistant bacterial infection, number of preoperative cardiac and cerebrovascular complications, functional status, transplant type, and operation time (h). The distribution diagram of LASSO regression coefficients and lambda decision diagram are attached. Figure 2 , 3 .

[0068] 6.3 Prediction Model

[0069] 6.3.1 Logistic Regression

[0070] The data set was randomly divided into training set and test set in a ratio of 7:3, and the LASSO regression screening variables were incorporated into the l2 regularized Logistic. The l2 regularization algorithm reduces the complexity and overfitting risk of the Logistic regression model by adding the sum of the squares of a parameter multiplied by a regularization coefficient in the loss function, thereby ultimately determining the optimal model parameters, where the l2 regularization coefficient is 1.0.

[0071] The results showed that the optimal model established by l2 regularized Logistic regression had a training set AUC of 0.745 (0.713-0.778) and a test set AUC of 0.670 (0.614-0.725).

[0072] 6.3.2 Random Forest

[0073] The training data set is the same as Logistic regression, and the variables included are also Lasso regression screening variables. The number of trees is 12, the maximum tree depth is 2, the maximum number of features is 5, and the variable selection standard uses the Gini coefficient, and then a 5-fold cross-validation is performed to finally determine the optimal model parameters.

[0074] The results showed that the optimal model established by random forest had a training set AUC of 0.741 (0.708-0.774) and a test set AUC of 0.677 (0.622-0.732).

[0075] 6.3.3 Deep Neural Networks

[0076] The training data set is the same as Logistic regression, and the variables included are also Lasso regression screening variables. Through a 2-layer fully connected neural network, each layer contains 50 neural networks, the dropout probability is 0.2, the activation function uses the LeakyReLU function, the optimizer uses the Adama strategy, the learning rate is 1.00×-4, and the cross entropy loss is used as the model optimization target to finally determine the optimal model parameters.

[0077] The results showed that the optimal model established by the deep neural network had a training set AUC of 0.865 (0.840-0.889) and a test set AUC of 0.779 (0.731-0.827).

[0078] 6.3.4 Convolutional Neural Networks

[0079] The training data set is the same as Logistic regression, and the variables included are also Lasso regression screening variables. Through a 6-layer neural network, including 2 convolutional layers, 2 maximum pooling layers, 1 dimension transformation layer, and 1 fully connected layer, the activation function uses the LeakyReLU function, the optimizer uses the Adama strategy, the learning rate is 1.00×-4, and the cross entropy loss is used as the model optimization target to finally determine the optimal model parameters.

[0080] The results showed that the optimal model established by the convolutional neural network had a training set AUC of 0.949 (0.936-0.962) and a test set AUC of 0.855 (0.816-0.894).

[0081] 7. Model comparison

[0082] Delongtest was used to compare the ROC curves of the training set and the test machine among different models. The McNemar test was used to test and compare the discrimination ability of each model. The performance of the models was compared and analyzed from the perspective of Brier score and decision curve.

[0083] 7.1 Delong test

[0084] By performing Delong test on the ROC curves of each model, the ROC curves of the main comparison models in the training set and the test set are shown in the attached figure. Figure 7 , 8 .

[0085] The results showed that there was no statistically significant difference between the ROC curves of the training set and the test set for Logistic regression and random forest. In addition, the differences between the ROC curves of other models were statistically significant. According to the AUC performance, the CNN model performed best in terms of ROC curve (training set AUC = 0.949 (0.936-0.962), test set AUC = 0.855 (0.816-0.894)), followed by MLP (training set AUC = 0.865 (0.840-0.889), test set AUC = 0.779 (0.731-0.827)).

[0086] 7.2 McNemar test

[0087] The McNemar test was used to analyze the classification ability of the models.

[0088] The results show that there is no statistically significant difference in the classification ability of Logistic regression and random forest between the training set and the test set. In addition, the classification ability of other models is statistically significant. According to the confusion matrix of each model, the CNN model has the best classification ability, followed by the MLP.

[0089] 7.3Brier Rating

[0090] Brier score is often used to measure model classification performance. The smaller the value, the stronger the predictive ability. Usually, a value between 0 and 0.1 is considered to have excellent predictive ability, a value between 0.1 and 0.25 is considered to have good predictive ability, and a value above 0.25 indicates that the predictive ability of the model is average. The higher the score, the worse the predictive ability.

[0091] Combined with the Brier score results of each model, the prediction ability of each model is good.

[0092] 7.4 Decision Curve

[0093] The decision curve is used to analyze and evaluate the net benefits of the prediction model. According to the comparison of the decision curves of each model, the deep convolutional neural network and deep neural network are generally higher than the logistic regression and random survival forest models in terms of net benefits. See the attached Figure 6 .

[0094] Methodological description

[0095] All statistical analyses were completed using R version 4.2.2 (2022-10-31) and Python 3.9.12 (main, Apr 4 2022, 05:22:27) software. All statistical tests were two-sided, and P < 0.05 was considered statistically significant.

[0096] Inter-group comparison was completed using the Mengte Cloud Statistics (https: / / mengte.pro) tool, SMOTE upsampling was completed using the UBL (version0.09) R package, LASSO regression was completed using the glmnet (version4.1.8) R package, L2 regularized logistic regression and random forest were completed using scikit-learn (version1.3.0), and deep neural networks and convolutional neural networks were completed using Pytorch (version2.2.0+cu121).

[0097] Intergroup comparison: The measurement data that obeyed the normal distribution were expressed as mean ± standard deviation (Mean ± SD), and the independent sample t test was used for intergroup comparison. The measurement data that did not obey the normal distribution were expressed as median and interquartile range [M(Q1, Q3)], and the MannWhitneyU test was used for intergroup comparison. The count data were expressed as number of cases and constituent ratio [n(%)], and the Pearsonχ test was used for intergroup comparison of unordered categorical data. 2 Mann-Whitney U test was used for comparison of ordered categorical data between groups.

[0098] Dataset interpolation: Some variables have missing values, so simple interpolation is used, i.e., continuous variables are interpolated by mean, and categorical variables are interpolated by mode. The description of data before interpolation and the comparison of components before and after interpolation are detailed in the attached table.

[0099] Data upsampling: The initial data set had 717 records. The ratio of the infected group and the non-infected group of the dependent variable "clinical bacterial infection" was 591 / 126. The ratio deviation was large. The smote upsampling algorithm was used to sample the ratio of the two groups of variables to a ratio of 591 / 592, and the final sample size was 1,183 cases.

[0100] Variable screening: With clinical bacterial infection as the dependent variable, Lasso regression was used to screen the independent variables. The optimal lambda value was determined through cross-validation to screen the variables for inclusion in the analysis. Finally, 16 variables were successfully screened out of 78 variables. For details, see the Lasso analysis section.

[0101] Prediction model: The total data set is 1,183 cases. The data set is first randomly divided into training set and test set with a ratio of 70:30. The training set has 828 cases and the test set has 355 cases. The prediction model is trained with "clinical bacterial infection" as the dependent variable. The models use l2 regularized logistic regression, random survival forest, deep neural network and convolutional neural network respectively.

[0102] Model evaluation and comparison: ROC curve, sensitivity, specificity, Brier score, decision curve and decision score are used to evaluate the predictive ability of each model. Delongtest is used to compare the ROC curves of each model, McNemar test is used to test the discrimination ability of each model, and decision curve comparison is used to compare and analyze the models.

[0103] The descriptive analysis results are shown in the table below.

[0104] Table 1. Descriptive analysis

[0105]

[0106]

[0107]

[0108]

[0109] LASSO analysis

[0110] LASSO regression was used to screen the variables. Through 10-fold cross validation, the optimal lambda was finally determined to be 0.019, and 16 independent variables were screened out, namely: donor type, donor BMI, donor ABO blood type, cold ischemia time (h), major lung disease diagnosis classification, recipient age (years), recipient BMI, smoking, pulmonary embolism, lung infection, pulmonary sepsis requiring treatment 2 or more times in the past year, pan-resistant bacterial infection, number of preoperative cardiac and cerebral complications, functional status, transplant type, and operation time (h). The coefficients of each variable are detailed in the table below.

[0111] Table 2. LASSO regression coefficients

[0112]

[0113]

[0114] Prediction Model

[0115] Logistic

[0116] The data set was randomly divided into training and test sets in a ratio of 7:3. The LASSO regression screening variables were incorporated into the l2 regularized Logistic, where the l2 regularization coefficient was 1.0. The specific performance of the model is described as follows.

[0117] Table 3. Model coefficients

[0118]

[0119] Table 4. Logistic model evaluation indicators

[0120]

[0121] Random Forest

[0122] The training data set is the same as Logistic regression, and the variables included are also LASSO regression screening variables. After 5-fold cross validation, the optimal model parameters are finally determined. The main parameters include: the number of trees is 12, the maximum tree depth is 2, the maximum number of features is 5, and the variable selection standard uses the Gini coefficient. The specific performance of the model is detailed below.

[0123] Table 5. Importance of model variables

[0124]

[0125]

[0126]

[0127] Table 6. RandomForest model evaluation indicators

[0128]

[0129] Deep Neural Networks

[0130] The training data set is the same as Logistic regression, and the variables included are also LASSO regression screening variables. After 5-fold cross validation, the optimal model parameters are finally determined. The model includes 2 layers of fully connected neural networks, each layer contains 50 neural networks, the dropout probability is 0.2, the activation function uses the LeakyReLU function, the optimizer uses the Adama strategy, the learning rate is 1.00×-4, and the cross entropy loss is used as the model optimization target. The specific performance of the model is detailed below.

[0131] Table 7. MLP model evaluation indicators

[0132]

[0133] Convolutional Neural Networks

[0134] The training data set is the same as Logistic regression, and the variables included are also LASSO regression screening variables. After 5-fold cross validation, the optimal model parameters are finally determined. The model includes 6 layers of neural network, including 2 convolutional layers, 2 maximum pooling layers, 1 dimension transformation layer, and 1 fully connected layer. The activation function uses the LeakyReLU function, the optimizer uses the Adama strategy, the learning rate is 1.00×-4, and the cross entropy loss is used as the model optimization target. The specific performance of the model is detailed below.

[0135] Table 8. CNN model evaluation indicators

[0136]

[0137] Model comparison

[0138] Delongtest was used to compare the ROC curves of the training set and the test machine among different models. The McNemar test was used to test and compare the discrimination ability of each model. The performance of the models was compared and analyzed from the perspective of Brier score and decision curve.

[0139] Delong inspection

[0140] By performing Delong test on the ROC curves of each model, the ROC curves of the models in the training set and the test set are mainly compared.

[0141] The results show that there is no statistically significant difference between the ROC curves of Logistic regression and random forest in the training set and the test set. In addition, the differences between the ROC curves of other models are statistically significant. According to the AUC performance, the CNN model performs best in terms of ROC curve, followed by MLP.

[0142] Table 9. Delong test of training set

[0143]

[0144] Table 10. Delong test of test set

[0145]

[0146] McNemar's test

[0147] The McNemar test was used to analyze the classification ability of the models.

[0148] The results show that there is no statistically significant difference in the classification ability of Logistic regression and random forest between the training set and the test set. In addition, the classification ability of other models is statistically significant. According to the confusion matrix of each model, the CNN model has the best classification ability, followed by the MLP.

[0149] Table 11. McNemar test of training set

[0150]

[0151] Table 12. McNemar test on test set

[0152]

[0153]

[0154] Brier score

[0155] Brier score is often used to measure model classification performance. The smaller the value, the stronger the prediction ability. Generally, the model prediction ability is excellent when it is between 0 and 0.1, good when it is between 0.1 and 0.25, and average when it is above 0.25. The higher the score, the worse the prediction ability. The Brier score of each model is shown in the table below.

[0156] Combined with the Brier score results of each model, the prediction ability of each model is good.

[0157] Table 13. Model Brier score

[0158]

[0159] Decision Curve

[0160] According to the comparison of the decision curves of each model, deep convolutional neural network and deep neural network are generally higher than logistic regression and random survival forest models in terms of net benefits.

[0161] in conclusion

[0162] Four models, namely l2 regularized logistic regression, random survival forest, deep neural network and deep convolutional neural network, were run to establish prediction models respectively. The models were evaluated using indicators such as sensitivity, specificity, accuracy, AUC and NetBenefits. The Delong test, McNemar test, Brier score and decision curve were used to compare the performance of each model horizontally.

[0163] After model evaluation and comparative analysis, the convolutional neural network performed best in terms of ROC curve, AUC area, classification ability, etc., followed by the deep neural network, and the logistic and random forest models performed equally well. In terms of maximum clinical benefit, the net benefits of the convolutional neural network and deep neural network were higher than the other two models.

[0164] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a prediction model for clinical bacterial infection factors in lung transplant patients, characterized in that: The following steps are involved: S1: Determine the research subjects: Select the hospitals to be studied and determine the research subjects according to the inclusion and exclusion criteria; S2: Selection of candidate predictors: Select an appropriate number of variables as candidate predictors; S3: Definition of outcome indicators: Patients with clinical bacterial infection are defined as: typical symptoms, signs and sputum characteristics; laboratory examination: elevated inflammatory indicators, sputum culture detection of bacteria to identify pathogens; imaging examination: pneumonia chest CT or X-ray examination shows new or progressive infiltration, consolidation or ground-glass shadows, bronchitis imaging examination is normal or chest X-ray shows moderate interstitial exudate; S4: Missing value processing: Some variables have missing values, which are processed by simple interpolation: continuous variables use mean interpolation, and categorical variables use mode interpolation; S5: Sample size analysis: First, the initial data set is obtained. When the ratio of the infected group and the non-infected group of the variable "clinical bacterial infection" is greatly deviated, the SMOTE upsampling algorithm is used to adjust the ratio of the two groups of variables to a certain ratio, and finally the sample size is obtained; S6: Model building: First, descriptive analysis is performed, followed by Lasso regression and prediction model.

2. The method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients according to claim 1, characterized in that: The inclusion criteria in the S1 step include the selection of lung transplantation surgery and unilateral or bilateral lung transplantation within a certain period of time, and the exclusion criteria include multi-organ combined transplantation, secondary transplantation, recipient age <18 years old and lack of complete medical records.

3. The method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients according to claim 1, characterized in that: The increased inflammatory indicators in the step S3 include increased white blood cells, neutrophils, C-reactive protein and serum procalcitonin.

4. The method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients according to claim 1, characterized in that: The analysis described in step S6 includes that the measurement data that follow normal distribution are expressed as mean ± standard deviation, and the independent sample t test is used for comparison between groups; The measurement data that did not conform to normal distribution were expressed as median and interquartile range, and the MannWhitney U test was used for comparison between groups; Enumeration data were expressed as number of cases and composition ratio. Pearson χ2 test or Fisher's exact probability method were used for comparison between groups of unordered categorical data, and MannWhitney U test was used for comparison between groups of ordered categorical data. Univariate analysis showed that 17 variables including age at diagnosis, BMI, and WHO had p values ​​less than 0.1, and these variables were included in subsequent model analyses.

5. The method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients according to claim 1, characterized in that: The LASSO regression in step S6 is a linear regression method that performs feature selection and parameter estimation by adding an L1 regularization term. The formula is: minimize1 / 2 * ||Y-Xβ||^2+λ*||β|| Among them, Y is the dependent variable vector, X is the independent variable matrix, β is the regression coefficient vector, ||.|| represents the norm, ||Y-Xβ||^2 represents the residual sum of squares, that is, the sum of the squares of the differences between the predicted values ​​and the actual observed values, and ||β|| is the L1 norm, which represents the sum of the absolute values ​​of the regression coefficient vector.

6. The method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients according to claim 1, characterized in that: The prediction models in step S6 include Logistic regression, random forest, deep neural network and convolutional neural network; The Logistic regression randomly divides the data set into a training set and a test set in a ratio of 7:3, and incorporates the LASSO regression screening variables into the l2 regularized Logistic. The l2 regularization algorithm reduces the complexity and overfitting risk of the Logistic regression model by adding the sum of the squares of a parameter multiplied by a regularization coefficient in the loss function, thereby ultimately determining the optimal model parameters, where the l2 regularization coefficient is 1.

0.

7. The method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients according to claim 6, characterized in that: The random forest uses the same training data set as Logistic regression, and the variables included are also Lasso regression screening variables. The variable selection standard uses the Gini coefficient and then undergoes 5-fold cross validation to finally determine the optimal model parameters.

8. The method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients according to claim 6, characterized in that: The deep neural network uses the same training data set as Logistic regression, and the variables included are also Lasso regression screening variables. Then, it uses a fully connected neural network, uses the LeakyReLU function as the activation function, uses the Adama strategy as the optimizer, and uses the cross entropy loss as the model optimization target to finally determine the optimal model parameters.

9. The method for constructing a prediction model of clinical bacterial infection factors in lung transplant patients according to claim 6, characterized in that: The convolutional neural network uses the same training data set as Logistic regression, and the variables included are also Lasso regression screening variables. Then, through the neural network, the activation function uses the LeakyReLU function, the optimizer uses the Adama strategy, and the cross entropy loss is used as the model optimization target to finally determine the optimal model parameters.

Images